search

WongKinYiu / yolov7

Implementation of paper - YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors
GNU General Public License v3.0
13.44k stars 4.23k forks source link

Keypoints on a custom dataset look wrong #1511

Open Jaykob opened 1 year ago

Jaykob commented 1 year ago

Hi! I’m currently trying to train it for pose detection on human feet, using 3 keypoints. Training looks OK looking at the values but when I run inference with detect.py and my webcam, the bounding boxes look fine, but the keypoints don’t make sense at all. They are out of the bounding box mostly and they do correspond to some extent to movement of the foot, but it seems as there’s a conversion problem or something.

Can somebody confirm that it works for them or any idea what the problem could be in my case?

ruiz-manuel commented 1 year ago

I am trying something similar. Did the kpt loss decrease in the training? show your results.txt

nomaad42 commented 1 year ago

Hi guys! @Jaykob @ruiz-manuel

Would you mind sharing the process of preparing dataset, modifications you made in code in order to train on custom dataset?

That would be really awesome!

ruiz-manuel commented 1 year ago

Have putmy code on my repo. Works for any nkpt and nc

nomaad42 commented 1 year ago

Have putmy code on my repo. Works for any nkpt and nc

Thank you so much!